Atlas-Based Classification Algorithms for Identification of Informative Brain Regions in fMRI Data.

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Machine Learning in Neuroscience

Background:

• Multi-voxel pattern analysis (MVPA) enhances sensitivity in neuroimaging compared to univariate methods.
• Searchlight, a common MVPA approach, faces limitations due to potential overestimation of activation regions based on sphere size.
• Computational cost is a significant factor in large-scale neuroimaging analyses.

Purpose of the Study:

• To evaluate the efficacy of two Searchlight alternatives: atlas-based local averaging (ABLA) and Multi-Kernel Learning (MKL).
• To identify informative brain regions supporting specific mental operations.
• To assess method performance across varying levels of differential BOLD activation and diverse brain atlases.

Main Methods:

• Comparison of Searchlight with ABLA and MKL for MVPA.
• Evaluation in scenarios with large versus small differential BOLD activation.
• Utilized nine different brain atlases to assess the impact of parcellation strategies.

Main Results:

• Both ABLA and MKL successfully localized informative regions when differential activation was large, showing stability across atlases.
• The weights provided by ABLA and MKL indicated the directionality of univariate approaches.
• ABLA was uniquely capable of localizing informative regions in scenarios with small differential activation.

Conclusions:

• Atlas-based methods like ABLA are viable alternatives to Searchlight for MVPA.
• The choice of method should consider the specific classification task and expected effect size.
• ABLA demonstrates particular utility when subtle neural differences need to be detected.